Kronos: Treating K-Lines as Language — The First Open-Source Foundation Model for Financial Markets
Technical analysts look at candlestick charts and see head-and-shoulders tops, double bottoms, hammer lines — these "patterns" are essentially "vocabulary" that human eyes identify in price sequences.
If an AI looks at K-lines, what does it see? Kronos's answer: treat K-lines as a language and use language-model methods to understand them.
The Old Problem of Financial Time-Series Forecasting
Financial time-series forecasting is notoriously difficult. Traditional methods (ARIMA, LSTM) treat prices as continuous numbers, but financial data has three fatal characteristics:
1. High noise: most price fluctuation is noise, not signal 2. Non-stationarity: distributions drift over time; rules that work today fail tomorrow 3. Fat-tailed distributions: extreme events occur more often than a normal distribution predicts
General-purpose time-series foundation models (TSFMs) underperform on finance because they never learned financial data's "dialect." They are trained on relatively clean series like electricity load and traffic flow, and struggle with the violent noise of markets.
Kronos's Two-Stage Framework
Kronos (GitHub Trending +217⭐/day, accepted at AAAI 2026) is the first open-source foundation model pre-trained specifically for financial markets. Its core innovation is a two-stage framework:
Stage 1: A specialized tokenizer quantizes OHLCV into hierarchical discrete tokens
OHLCV (open/high/low/close/volume) is continuous, multi-dimensional data. Rather than simple uniform bucketing, Kronos uses hierarchical quantization — capturing price movements at different scales.
This step is crucial. Uniform bucketing loses the multi-scale character of price movement: a 1% move and a 5% move could land in the same bucket. Hierarchical quantization gives small and large moves their own "vocabulary."
Stage 2: Autoregressive Transformer pre-training
Autoregressive pre-training on hierarchical discrete tokens, just like GPT — except the "vocabulary" is K-line tokens instead of natural language.
The elegance: it invented no new architecture; it invented a new "language." The Transformer and the pre-training recipe are off-the-shelf; the innovation lies in how continuous financial data is "translated" into discrete tokens.
Data and Models
Training data
K-line data from 45+ global exchanges. This coverage means Kronos learned not just the "dialect" of US stocks or A-shares, but a "lingua franca" of global markets.
Model family
| Model | Params | Context length | Open-source | |---|---|---|---| | Kronos-mini | 4.1M | 2048 | ✅ | | Kronos-small | 24.7M | 512 | ✅ | | Kronos-base | 102.3M | 512 | ✅ | | Kronos-large | 499.2M | 512 | ❌ |
Note: the mini version has only 4.1M parameters with a 2k context — remarkably small. GPT-class models run into the hundreds of billions of parameters, yet Kronos-mini can forecast with 4.1M. This suggests the "language" of financial time series is far simpler than natural language: smaller vocabulary, fewer grammar rules, shorter context dependencies.
Academic validation
Accepted at AAAI 2026; arXiv: 2508.02739, with full comparative experiments and ablation studies.
Engineering Insight: X as Language
Kronos hints at a deeper trend — the "X as language" framework is expanding into more domains:
- Natural-language GPT: treating text sequences as language
- VLMs: discretizing image patches into tokens
- Speech models: discretizing sound waves into tokens
- Kronos: discretizing K-lines into tokens
- Uniform bucketing = treating all price moves at equal granularity — if 1% and 5% moves land in the same bucket, scale information is lost
- Hierarchical quantization = multiple granularities coexisting — coarse scale for trends, fine scale for fluctuations, each with its own vocabulary
- Natively multimodal scaling laws: different modalities have different "languages," and specialized tokenizers are key — Kronos's financial tokenizer confirms this
- Möbius RoPE: positional encoding matters for sequential data — Kronos's context-length design (mini 2k, others 512) reflects considerations about temporal dependency depth
- Octopus RNA editing: don't change the blueprint (raw OHLCV data); change the construction plan (tokenization) — Kronos's innovation is in data representation, not model architecture
The common thread: discretize continuous/raw data into tokens, then pre-train a Transformer. The Transformer's real power is not in natural language but in any sequence that can be discretized into tokens.
The Subtlety of Hierarchical Quantization
"Hierarchical quantization" sounds simple, but it is Kronos's core innovation. An analogy:
This is isomorphic to how humans read charts. Technical analysts don't just look at 1-minute candles; they simultaneously read daily, weekly, and monthly charts — different scales reveal different signals. Kronos's hierarchical quantization encodes this multi-scale view into the tokenizer.
Cross-Paper Consensus
Kronos's design resonates with recent work:
Conclusion
The trend Kronos points to is bigger than financial prediction itself — the "X as language" framework is expanding. Financial K-lines are just the beginning; future "protein languages," "climate languages," and "gene languages" may follow. Once any continuous data can be discretized into tokens, the Transformer's application boundary extends from natural language to all sequential data.
The first open-source foundation model for financial markets is not just a new tool for quant traders — it is another validation of the "X as language" paradigm.
---
Project: https://github.com/shiyu-coder/Kronos
Paper: https://arxiv.org/abs/2508.02739
Live Demo: https://shiyu-coder.github.io/Kronos-demo/
Models: HuggingFace NeoQuasar/Kronos-{mini,small,base}
License: open source
Academic: accepted at AAAI 2026